Predicting responders to prone positioning in mechanically ventilated patients with COVID-19 using machine learning

Area of Science:

• Critical Care Medicine
• Pulmonary Medicine
• Medical Informatics

Background:

• Prone positioning is a key treatment for mechanically ventilated COVID-19 patients.
• However, it is labor-intensive and carries risks.
• Identifying suitable candidates can optimize resource allocation.

Purpose of the Study:

• To develop a predictive model for successful prone positioning in critically ill COVID-19 patients.
• To identify clinical parameters that predict response to prone positioning.

Main Methods:

• Utilized machine learning algorithms (Logistic Regression, Random Forest, XGBoost, etc.) on data from 1142 COVID-19 ICU patients.
• Predicted outcomes included improvements in PaO2/FiO2 ratio, ventilatory ratio, compliance, or mechanical power after 4 hours.
• Models were trained on readily available clinical features.

Main Results:

• Machine learning models showed poor discrimination between responders and non-responders (AUC 0.62 for PaO2/FiO2).
• Feature importance was inconsistent across different outcome models.
• Previous response to proning and pre-proning PEEP levels did not significantly predict success.

Conclusions:

• Predicting prone positioning success in mechanically ventilated COVID-19 patients using electronic health record data is not currently feasible.
• A liberal approach to proning for all severe COVID-19 ARDS patients is justified.
• Previous proning outcomes should not influence the decision to pronate.